Re: find_busiest_group using lots of CPU

[Jens Axboe <jens.axboe@oracle.com>]

On Mon, Oct 05 2009, Peter Zijlstra wrote:
> On Wed, 2009-09-30 at 10:18 +0200, Jens Axboe wrote:
> > Hi,
> > 
> > I stuffed a few more SSDs into my text box. Running a simple workload
> > that just does streaming reads from 10 processes (throughput is around
> > 2.2GB/sec), find_busiest_group() is using > 10% of the CPU time. This is
> > a 64 thread box.
> > 
> > The top two profile entries are:
> > 
> >     10.86%      fio  [kernel]                [k] find_busiest_group
> >                 |          
> >                 |--99.91%-- thread_return
> >                 |          io_schedule
> >                 |          sys_io_getevents
> >                 |          system_call_fastpath
> >                 |          0x7f4b50b61604
> >                 |          |          
> >                 |           --100.00%-- td_io_getevents
> >                 |                     io_u_queued_complete
> >                 |                     thread_main
> >                 |                     run_threads
> >                 |                     main
> >                 |                     __libc_start_main
> >                  --0.09%-- [...]
> > 
> >      5.78%      fio  [kernel]                [k] cpumask_next_and
> >                 |          
> >                 |--67.21%-- thread_return
> >                 |          io_schedule
> >                 |          sys_io_getevents
> >                 |          system_call_fastpath
> >                 |          0x7f4b50b61604
> >                 |          |          
> >                 |           --100.00%-- td_io_getevents
> >                 |                     io_u_queued_complete
> >                 |                     thread_main
> >                 |                     run_threads
> >                 |                     main
> >                 |                     __libc_start_main
> >                 |          
> >                  --32.79%-- find_busiest_group
> >                            thread_return
> >                            io_schedule
> >                            sys_io_getevents
> >                            system_call_fastpath
> >                            0x7f4b50b61604
> >                            |          
> >                             --100.00%-- td_io_getevents
> >                                       io_u_queued_complete
> >                                       thread_main
> >                                       run_threads
> >                                       main
> >                                       __libc_start_main
> > 
> > This is with SCHED_DEBUG=y and SCHEDSTATS=y enabled, I just tried with
> > both disabled but that yields the same result (well actually worse, 22%
> > spent in there. dunno if that's normal "fluctuation"). GROUP_SCHED is
> > not set. This seems way excessive!
> 
> io_schedule() straight into find_busiest_group() leads me to think this
> could be SD_BALANCE_NEWIDLE, does something like:
> 
> for i in /proc/sys/kernel/sched_domain/cpu*/domain*/flags; 
> do 
> 	val=`cat $i`; echo $((val & ~0x02)) > $i; 
> done
> 
> [ assuming SCHED_DEBUG=y ]
> 
> Cure things?

I can try, as mentioned it doesn't look any better with SCHED_DEBUG=n

> If so, then its spending time looking for work, which there might not be
> on your machine, since everything is waiting for IO or somesuch.

OK, just seems way excessive for something which is only 10 tasks and
not even that context switch intensive.

> Not really sure what to do about it though, this is a quad socket
> nehalem, right? We could possibly disable SD_BALANCE_NEWIDLE on the NODE
> level, but that would again decrease throughput in things like kbuild.

Yes, it's a quad socket nehalem. I'll see if disabling NEWIDLE makes a
difference, I need to run some other tests on that box today anyway.

-- 
Jens Axboe